Meeting 5
Meeting 5
Get a virtual background.
Gather in groups of 3-5.
Open the Google Doc (link removed) for today.
Virtual Backgrounds
Each of these backgrounds illustrates one of the resources essential to the development of modern AI.
The historical development of modern computation
Diverse sources of training data
New forms of electrical power
Contemporary GPU chips
Activity: Why does size matter?
We know AI companies are spending huge amounts of money to get even bigger models trained with more computing power using more data and producing larger neural nets. But we also know that they keep making small models too. Today we will try to understand why.
Go to TextSynth to try out some text completion on different-sized models. These are just text-completion - they don’t have the specialized chatbot training, or reasoning, so what you see is a direct indication of the base model itself. Something like this is under the hood of the more familiar chatbots, but here it’s in pure form, which makes it a bit harder to use.
We will compare the “LLaMa 3.1 8B instruct” model against the “LLaMa 3.3 70B instruct” model - these are two versions of Meta’s LLaMa model, where the main difference is whether the LLM has roughly 400,000 neurons (and 8 billion weights and biases) or a few million neurons (and 70 billion weights and biases).
Try each prompt five times with each model. Keep track of these on the Google Doc (link removed) for today. Note how long the model takes to give each response (“right away”, or if it takes 2 or 3 seconds, or maybe even try to time it if it’s taking longer). Also, discuss with each other which responses are better or worse. You might also note that it does some surprising things when completing the text. What can you figure out about what kinds of documents the LLaMa instruct systems were probably trained on?
If anyone finds a particularly interesting or surprising response, note that at the top of your group’s page.
Sample prompts:
What is the capital of France?
USER: What is the capital of France? ASSISTANT:
The capital of France is
If you don’t get enough vitamin C, you might get scurvy. If you don’t get enough iron, you might get anemia. If you don’t get enough vitamin B12, you might
Irvine, Calif., Apr. 28, 2025. Researchers at the University of California have developed a new form of energy storage far more efficient than lithium ion batteries. The new technology
At dawn, my dreams are torn
Looking through the cracks
I see a painting
After doing this for a while, you might want to try chatbots rather than text-completion, go to ZotGPT chat (sign in with your UCI id) and select different models. (You’ll have to start a new chat to select each one.)
- For Anthropic, the biggest and most powerful is Claude Opus 4.8 (under “For Advanced Users”) while the oldest and smallest is Claude Sonnet 3.5v2 (under “Legacy”)
- For Google, the biggest and most powerful is Google Gemini 3.5 Flash (under “For Advanced Users”) while the oldest and smallest is Google Gemini 2.5 Flash-Lite (under “Legacy”)
- For OpenAI, the biggest and most powerful is GPT-5.5 (under “Recommended”), while the oldest and smallest is GPT-4.1 Nano (under “Legacy”)
- You can also find some interesting ones from other companies: Deepseek R1 (“For Advanced Users”), Kimi K2.5 (“Other”), and three versions of Mistral: Large 3 (“For Advanced Users”), 8x7B Instruct (“Other”), Large 2 (“Legacy”)
For the chatbot versions, you might want to try some of the prompts from Week 2, or other things you’ve been thinking about lately.
At a certain point, you might look through the document to see what’s going on with other groups.
Second activity: Discuss this module’s ideas
Open up the Canvas page for module 5 (link removed) to remind yourselves of the videos for this module. Discuss these videos with each other.
Were there any ideas from any of them that you found particularly surprising, or interesting? Were there things you didn’t understand?
Look at the list of “Further watching/reading” links below the actual assignments. Are there some that look interesting to dig into further? Are there any you already looked at that some of the other people in your group might want to check out?